On Tue, 26 Jan 2021 11:41:52 GMT, Aleksey Shipilev <sh...@openjdk.org> wrote:
> I was following up on performance testing results for some ongoing work, and > realized that there are `ShenandoahMarkContext::mark_strong` calls from > `mark_work_loop`. Those callees were supposed to be inlined. I believe it is > a simple omission after JDK-8255019 moved `mark_work_loop` code to > `shenandoahMark.cpp`. > > On a typical workload: > > # Before > [44.500s][info][gc,stats] Concurrent Marking = 0.395 s (a = 11278 us) > (n = 35) (lvls, us = 6426, 9473, 11133, 12891, 16476) > > # After > [44.405s][info][gc,stats] Concurrent Marking = 0.337 s (a = 9636 us) > (n = 35) (lvls, us = 3770, 7383, 9785, 11328, 16776) > > Additional testing: > - [x] Eyeballing GC hot paths > - [x] Ad-hoc performance tests Looks good from the build PoV. ------------- Marked as reviewed by ihse (Reviewer). PR: https://git.openjdk.java.net/jdk/pull/2235
